MiniMax First Half Report: ARR Exceeds $800M | Luzhou Shengmingli

On August 26, Oasis Capital portfolio company MiniMax released its first mid-year report since going public: revenue of approximately $120 million in the first half of 2026, up 283% year-over-year; August ARR has already exceeded $800 million.

This report spotlights the company's growth trajectory and sends a clear signal: as MiniMax continues to improve its model capabilities, it is crossing the critical threshold for scaled commercial deployment. We have always believed in the team's long-term commitment to AGI, and we will continue to stand by them as their vision of "Intelligence with Everyone" gradually becomes reality.

Below is media coverage of the mid-year report. Enjoy.

On August 26, 2026, MiniMax delivered its first-half results — the first mid-year report among large model companies. For the past two years, the industry talked mostly about parameters, leaderboards, and fundraising. But the question on everyone's mind now has shifted to "Can ever-increasing compute investment translate into real usage and sustainable revenue?" MiniMax's first report offers a rare public data point.

Let's start with the most striking figures: MiniMax's revenue and usage are surging in tandem. August ARR has already surpassed $800 million, while July token consumption reached 20x January levels. The revenue mix has also shifted dramatically: B2B revenue grew more than 7x year-over-year and now accounts for roughly 80% of total revenue, compared to about 30% in the same period last year — nearly a complete inversion of the B2B/B2C structure.

MiniMax also addressed ARR growth, intelligence improvements, inference efficiency, compute reserves, and other key topics during its earnings call. Here's our summary of the most critical takeaways.

August ARR rises further to over $800 million

Overseas revenue accounts for ~60%

H1 2026 revenue reached approximately $120 million, up 283% year-over-year — 1.5x full-year 2025 revenue in just six months. Q2 revenue grew 81.8% quarter-over-quarter. July token consumption hit 20x January levels. August ARR climbed further to over $800 million.

B2B customer scale exploded and revenue mix improved significantly: In H1 2026, B2B (open platform and other AI enterprise services) revenue was approximately $74 million, up 7x year-over-year, with its share of total revenue rising from 30% in the prior year to 63%. In August ARR, B2B contribution already exceeded 80% (versus 30% B2B / 70% B2C last year). Enterprise customers and developers surpassed 2 million, 10x year-end 2025 levels, with clear acceleration in enterprise and developer adoption.

Overseas revenue leads. In H1 2026, overseas revenue accounted for approximately 60%, domestic for roughly 40%.

H1 gross profit was approximately $21 million, up 465% year-over-year. The company expects gross margins to continue improving in H2 2026, with further room for improvement next year.

The core driver of ARR growth is model capability improvement. The company also observes that model consumption is expanding from "human-AI interaction" to "multi-turn agent-AI interaction," with agent-driven inference demand growing significantly faster than human user counts or message volumes. Per-user token consumption is rising rapidly.

From Stronger Models to Cheaper Intelligence:

MiniMax's Technical Roadmap and Business Thesis

MiniMax founder and CEO Junjie Yan said on the call: "What MiniMax pursues is not a trade-off between intelligence and cost. Only by minimizing the cost per unit of intelligence can we train the highest level of intelligence; only by reducing the supply cost per unit of intelligence can higher-level intelligence reach broader production and daily life."

He emphasized that compute is a constraint for every company. "Minimize the Inference Cost, Maximize the Intelligence" to achieve "Intelligence with Everyone" is the technical path we adhere to and our original mission as founders.

There remains substantial room for model intelligence improvement: Technically, models still face challenges in long-horizon planning, tool use, result verification, cross-modal understanding, and stability on professional tasks. The Scaling dimension itself is expanding — beyond Pre-training, Mid-training, Post-training, reinforcement learning, inference-time compute, synthetic data, evaluation, and agent environments are all emerging as new Scaling directions, with industry exploration still in early stages.

The next wave of scaled applications after coding: Multimodal productivity (complete workflows in advertising, gaming, e-commerce, design, and content production — moving from generation to understanding, modification, and final delivery), as well as high-value tasks with professional constraints and feedback loops such as cybersecurity and chip design. Business models may also evolve from token-based pricing toward task-outcome and professional-value-based pricing.

When model capabilities, usage costs, and product experience cross certain thresholds, real demand is released in step-function jumps rather than simple linear growth.

From Compute to Foundation Model:

MiniMax Continues Expanding Scaling Boundaries

As one of the first two independent large model companies in China to build scaled, long-term stable infrastructure, MiniMax's infrastructure can now independently support its largest training jobs with ETTR (Effective Training Time Ratio) reaching 97%. Based on full-stack infrastructure capabilities, there remains at least 3x room to improve post-training scale per unit of capital efficiency.

Compute reserves are built through a composite supply network combining proprietary core clusters, cloud provider supply, and Token Factory partnerships. M3 and H3 are both advancing domestic chip adaptation; the company will soon bring large-scale domestic compute clusters online to gradually handle real production traffic.

M3.1 and M3 Pro are progressing as planned. M3.1 builds on M3 with extensive pre-training and continued scaling of post-training, positioned as a model for mass, widespread use, with focused improvements in task completion quality, stability, inference efficiency, and agent environment compatibility. M3 Pro will scale parameters to approximately 3T, continuing to expand reinforcement learning and long-horizon task training on a larger base model, with stronger generalization capabilities and higher intelligence ceiling.

H3 Validates a New Path

Language Models Are Reshaping Visual Generation

Unlike most companies, MiniMax has designed models from the ground up with native multimodal considerations, believing that visual understanding and generation are essential components of productivity. Multimodal generation is currently the second-largest market for AGI after programming.

H3 is a practical implementation combining language models with visual generation. This goes beyond simply using language models as encoders — more importantly, it leverages language models for fine-grained contextual awareness, enabling comprehensive reference-based generation with precise control.

H3 has disrupted the pattern where state-of-the-art models belonged exclusively to major tech companies following closed approaches. Since H3's open-source release three weeks ago, it has been downloaded over 24 million times, spawned more than 300 public derivative models, and become one of the most downloaded models globally this year. Official online services have also seen explosive growth in usage.

H3 validates two things: first, language model capabilities can further raise the ceiling of visual generation models; second, open models, products, and APIs can form a mutually reinforcing ecosystem rather than cannibalizing each other. H3 brings language models into the creative process, understanding context, reference materials, and user intent, with clear improvements in subject consistency, reference image adherence, shot control, audio coordination, and iterative refinement.

Looking further ahead, MiniMax believes similar paradigms combining language and visual generation models will continue to emerge in more frontier domains. MiniMax believes that deep fusion of multimodality with language models will become an important direction driving the next leap in intelligence.